The paper introduces BELT (BELief-based Termination), a method that decouples outcome prediction from stopping control in large language models (LLMs). BELT learns beliefs about current and future answers from a frozen LLM’s hidden states and applies a linear stopping policy, enabling a single probe to adapt to various token‑price trade‑offs without retraining. Across five out‑of‑distribution math benchmarks, BELT cuts token usage by up to 51% while maintaining accuracy, and its lightweight probe outperforms stronger baselines in both efficiency and parameter count, also transferring effectively to coding and agentic tasks.
arXiv:2606. 30852v1 Announce Type: new Abstract: Reasoning models spend different amounts of useful computation across instances, but it remains unclear when a learned stopping rule improves over simple confidence or convergence thresholds.
arXiv:2609.16055v1 Announce Type: cross Abstract: Test-time compute has emerged as a major approach to improving the capabilities of Large Language Models (LLMs). However, existing test-time reasonin...
arXiv:2608. 20256v1 Announce Type: new Abstract: Reasoning language models trained with reinforcement learning typically operate under a fixed token budget rather than an explicitly adaptive one, which can lead to over-computation on easy problems and insufficient computation on difficult ones.
arXiv:2609.37304v1 Announce Type: new Abstract: Large reasoning models improve performance on challenging problems by allocating additional computation before answering, but longer reasoning does not...
The paper introduces PORA, a method that teaches large language model agents to decide when to allocate more reasoning steps versus fewer, based on the expected downstream benefit. PORA trains by running each state twice—once with short reasoning and once with long reasoning—and learns to predict the value of extended reasoning. During inference, a controller uses this prediction and remaining budget information to choose the appropriate reasoning length without running both options, improving performance across diverse tasks.
The paper introduces StreamStop, a lightweight stopping policy for reasoning-enabled language model agents that learns offline from prefix-level action agreement. By encoding available reasoning text with a frozen prefix reader and a stop head, StreamStop can decide when to stop reasoning without needing internal signals from the model. Experiments on WebShop and ScienceWorld show that StreamStop achieves higher success rates than many fixed-budget baselines while using fewer reasoning tokens, notably reducing token usage by 42% on WebShop with Qwen3.8-27B.
arXiv:2606. 19808v1 Announce Type: new Abstract: Test-time reasoning is increasingly used as a serving-time control knob, but extra reasoning is not uniformly valuable: it can repair failed attempts, waste compute on already-correct answers, or introduce harmful answer changes.
arXiv:2605.11467v2 Announce Type: replace-cross Abstract: Reasoning models post-hoc rationalize answers they have already committed to internally, producing chains of *reasoning theater*: deliberativ...
The paper introduces a self‑distillation method that leverages answer‑only data to improve large language models (LLMs) after training. By training an inverse reasoning LLM to generate latent chain‑of‑thought traces conditioned on the answer, the authors can perform offline distillation that significantly boosts performance on hard prompts for tasks like Countdown and GSM8K. Additionally, the inverse model serves as an accurate verifier, enhancing best‑of‑n selection accuracy.
The paper introduces EvoResearcher, a training‑free, inference‑time protocol that enables a frozen large language model to perform cost‑bounded self‑reflection and early stopping. By iterating through generate → self‑critique → revise steps until a maximum depth or a CONFIRMED sentinel is reached, the model can self‑verify its answers within a strict compute budget. The protocol incorporates four self‑reflective meta‑reward components—correctness, efficiency, reflection depth, and tool‑call diversity—implemented as prompt‑level mechanisms, and is validated on Big‑Bench Hard, GSM8K, and MATH benchmarks, achieving comparable accuracy while terminating 82‑88% of items early with only about 2.1 generations per question.
The paper introduces Trimmed Logit-Gap SFT (TrimSFT), a token-level reweighting strategy that adjusts supervised fine-tuning loss based on the logit gap between the correct token and its strongest competitor. TrimSFT trims supervision from tokens that are either already mastered (large logit gap) or poorly supported (small or negative logit gap), focusing learning on tokens with intermediate logit gaps. Experiments on six base models across five mathematical reasoning benchmarks show that TrimSFT consistently outperforms standard SFT, achieving the best average performance on five of six models and up to +26.9 points on MATH500.
arXiv:2608. 04771v1 Announce Type: new Abstract: Large Reasoning Models (LRMs) excel on complex tasks through long chain-of-thought (CoT) reasoning, but their lengthy intermediate steps cause severe overthinking that inflates inference cost.